Munich Personal RePEc Archive Stock Returns and Investors’ Mood: Good Day Sunshine or Spurious Correlation? Kim, Jae 12 April 2016 Online at https://mpra.ub.uni-muenchen.de/70692/ MPRA Paper No. 70692, posted 15 Apr 2016 07:00 UTC Stock Returns and Investors’ Mood: Good Day Sunshine or Spurious Correlation? Jae H. Kim Department of Economics and Finance La Trobe University, VIC 3086 Australia Abstract This paper examines the validity of statistical significance reported in the seminal studies of the weather effect on stock return. It is found that their research design is statistically flawed and seriously biased against the null hypothesis of no effect. This, coupled with the test statistics inflated by massive sample sizes, strongly suggests that the statistical significance is spurious as an outcome of Type I error. The alternatives to the p-value criterion for statistical significance soundly support the null hypothesis of no weather effect. As an application, the effect of daily sunspot numbers on stock return is examined. Under the same research design as that of a seminal study, the number of sunspots is found to be highly statistically significant although its economic impact on stock return is negligible. Keywords: Anomaly, Behavioural finance, Data mining, Market efficiency, Sunspot numbers, Type I error, Weather. JEL Classification: G12, G14. April 2016 Tel: +613 9479 6616; Email address: [email protected] I would like to thank Xiangkang Yin for helpful comments on an earlier version of the paper. 1 1. Introduction The question as to whether investors’ mood affects the stock market (i.e. their emotional states or feelings unrelated to market fundamentals or rational pricing of financial assets) has been an issue of considerable interest in economics and finance (see, for a survey, Lucey and Dowling, 2005). The seminal studies in this literature are Saunders (1993) and Hirshleifer and Shumway (2003) where statistically significant weather effect on stock return is reported. Subsequent studies overall support the existence of the weather effect (cloudiness, sunshine, temperature, or wind) on stock return or other trading activities: see Cao and Wei (2005), Dowling and Lucey (2005, 2008), Goetzmann and Zhu (2005), Chang et al. (2008), Chang et al. (2006), Keef and Roush (2002, 2005, 2007), Yoon and Kang (2009), Kang et al. (2010), Lee and Wang (2011), Lu and Chou (2012), and Novy-Marx (2014). The literature has proliferated over the years in the publication of studies examining the effects of investors’ moods derived from disparate sources such as: daylight saving (Kamstra et al. 2000), seasonal depression (Kamstra et al., 2003), sport events (Edmans et al., 2007; Kaplanski and Levi, 2010; Chang et al., 2012), lunar phases (Yuan et al., 2006; Keef and Khaled, 2012), pollution (Lepori, 2016), and the Ramadan (Bialkowski et al., 2012). Most of these studies report a statistically significant effect of investors’ mood on the stock market, and their findings are presented as direct evidence for the anomalies against market efficiency. On the other hand, there are studies that raise suspicions that a statistically significant weather effect may be the result of data mining or spurious correlation. In replicating Saunders’ (1993) results using a German data set, Krämer and Runde (1997) report that statistical significance of the weather effect depends largely on how the null hypothesis is phrased. Trombley (1997) provides evidence that Saunders’ (1993) results depend on the type of the return used and sample period employed. Loughran and Schultz (2004), in the context of localized trading of NASDAQ stocks, examine the weather effect in the city where the company is based and find 2 that the weather effect is too slight to establish a profitable weather-based trading strategy. They (p.363) state that “we would not dismiss the possibility that the relationship between cloud cover in New York and stock returns is spurious”. Jacobsen and Marquering (2008) state that the documented weather effects might be the consequence of “data-driven inference based on spurious correlation”, providing evidence that seasonal anomaly in stock return is unlikely to be caused by investors’ mood changes due to weather variations. The purpose of this study is to examine the validity of statistical significance reported in the two seminal studies of the weather effect on stock return, i.e. Saunders (1993) and Hirshleifer and Shumway (2003), in order to shed light on the possibility of a spurious relationship between investor’s mood and stock return. First, paying attention to Hirshleifer and Shumway (2003), I evaluate whether their research design that “maximizes the power of the test by pooling the all available data jointly” is statistically sensible. This is important since many subsequent studies in this area (and also elsewhere in finance) adopt large or massive sample sizes in the same spirit. However, there is a danger that the use of massive sample size produces spurious statistical significance (see, for example, McCloskey and Ziliak, 1996). Second, statistical significance reported in these seminal studies is re-evaluated using the Bayesian method (Zellner and Siow, 1978) and the adaptive level of significance (Perez and Pericchi, 2014). These are the alternatives to the p-value criterion for statistical significance exclusively adopted in prior studies. Note that the American Statistical Association (Wasserstein and Lazar, 2016) recently issued a statement expressing grave concerns that improper use of the p-value criterion is distorting the scientific process and invalidating many scientific conclusions. Third, as an application, the effect of sunspot numbers on stock return is examined, using the same research design as that of Hirshleifer and Shumway (2003). This is to demonstrate that a variable with little economic relevance on stock return can be shown to be statistically significant through a simple data mining process. 3 The main finding of the paper is that statistically significant weather effect reported in the past studies is highly likely to be spurious and an outcome of Type I error. In particular, the research design employed by Hirshleifer and Shumway (2003) is problematic, with the probability of Type I error disproportionately higher than that of Type II. It is severely biased against the null hypothesis of no effect in its implied specification of loss function and prior probabilities. The alternatives to the p-value criterion show an overwhelming support for the null hypothesis of no weather effect. It is also demonstrated that a balanced research design in the context of the data employed by Hirshleifer and Shumway (2003) requires the sample size only under 2000. The results from the empirical application further confirm these findings. While suspicions concerning spurious statistical significance of the weather effect on stock return have been raised previously, this study is the first to assess the underlying statistical issues in the research design of the seminal studies and re-evaluates statistical significance of their results. In the next section, the research design of Hirshleifer and Shumway (2003) is examined. Section 3 presents further analyses based on the alternative criteria for statistical significance; and discussion on the effect size estimates reported in the past seminal studies. Section 4 presents the empirical application, and Section 5 concludes the paper. 2. Issues related with the Research Design In this section, the research design of Hirshleifer and Shumway (2003) is discussed with reference to the possibility of spurious statistical significance. The weather effect is typically tested in the regression model of the form: Yt = 0 + 1X1t + … + KXKt + ut, (1) where Y is the stock return, X1 is a weather variable (e.g. cloud cover), and other X’s represent the possible control variables. Hirshleifer and Shumway (2003) also consider the logit model where Y is an indicator variable. Under H0: 1 = 0, the weather has no effect on the stock 4 b1 return. The t-test statistic can be written as t , where b1 is the least-square estimator s / n for 1, n is the sample size, and s /n denotes the standard error of b1. Throughout the paper, following convention, the regression parameters are denoted as i’s; and the probability of Type I and II errors as and β, respectively (the power 1- β). 2.1 Background One common feature of the studies of the weather effect is the use of large or massive sample sizes: a survey of twenty papers in the literature finds that the average sample size used is around 6000 with the maximum being 92808. In addition, they conduct their statistical tests almost exclusively at the conventional level of significance such as 0.05. A number of authors warn that spurious statistical significance may occur in this scenario (Neal, 1987, p. 524; Connolly, 1989, p. 139; McCloskey and Ziliak 1996; p.102). However, many researchers seem to believe that a large or massive sample size is necessarily a desirable feature of a research design, delivering strong power to their statistical tests. Hirshleifer and Shumway (2003, p.1014) justify their use of a massive panel data set, asserting that “the panel increases our power to detect an effect. … Given high variability of returns, it is useful to maximize power by using a large number of markets”. However, as we shall see, this can cause statistical inference severely biased towards Type I error. The extreme power leads to an acute imbalance between the and , if a conventional level of significance is maintained. For example, suppose an extreme power (1-) of 0.99999 is achieved by pooling a massive panel data set. If the researcher conducts a test at the 5% level (= 0.05), the Type I error is 5000 times more likely to occur than the Type II error. As a result, if an error occurs, it is highly likely to be that of Type I, rejecting the true null hypothesis of no effect.